How else can an agent be evaluated besides LLM-as-a-Judge: what criteria can be used to understand if it works correctly or close to correct?